Head to head · 100 clips

Cleanvoice vs ElevenLabs Voice Isolator

On 100 noisy clips from Dolby's public UNIVERSE set, Cleanvoice Studio Sound 4.1 made 15.5% word errors against 23.8% for ElevenLabs Voice Isolator, and dropped 3.0 words per 100 against 7.1. It also changed fewer words (11.2 vs 15.3 per 100) and scored higher on speech quality (SCOREQ 4.54 vs 3.72) and speaker similarity (0.955 vs 0.928).

Last updated: 2 November 2026 · 100 clips from Dolby’s public UNIVERSE set · Full benchmark

Try Cleanvoice Studio Sound free See every metric

Word error rate (lower is better)
15.5%
Cleanvoice, vs 23.8% for ElevenLabs
Dropped words (lower is better)
3.0%
Cleanvoice, vs 7.1% for ElevenLabs
Speech quality (SCOREQ) (higher is better)
4.54
Cleanvoice, vs 3.72 for ElevenLabs

Head to head on every metric

ElevenLabs Voice Isolator against Cleanvoice Studio Sound 4.1, with the unprocessed input for context. Same 100 clips, run on 1 November 2026: Cleanvoice processed each clip separately; the other tools processed the 100 clips joined into one file. Method and all eight tools.

Cleanvoice Studio Sound 4.1 vs ElevenLabs Voice Isolator on 100 noisy clips from Dolby’s UNIVERSE validation set, 1 November 2026
Metric	Unprocessed input	ElevenLabs	Cleanvoice Studio Sound 4.1	Better
Word error rate ↓Omnilingual ASR on the output vs on the clean original	17.0%	23.8%	15.5%	Cleanvoice
Dropped words ↓Deletions per 100 reference words	5.9%	7.1%	3.0%	Cleanvoice
Changed words ↓Substitutions per 100 reference words	10.8%	15.3%	11.2%	Cleanvoice
Speech quality (SCOREQ) ↑Natural speech quality, 1 to 5	2.64	3.72	4.54	Cleanvoice
Overall quality (DNSMOS OVRL) ↑DNSMOS P.835 overall score, 1 to 5	2.27	3.11	3.35	Cleanvoice
Speaker similarity ↑How close the voice stays to the clean original	0.916	0.928	0.955	Cleanvoice
Intelligibility (ESTOI) ↑Against the clean reference	0.714	0.754	0.840	Cleanvoice
PESQ ↑Against the clean reference	1.55	1.78	1.88	Cleanvoice
Bandwidth ↑Median output bandwidth; the input stops at 8 kHz	8.0 kHz	11.8 kHz	17.3 kHz	Cleanvoice

Cleanvoice has the better number on every metric, though PESQ is close (1.88 vs 1.78). The ElevenLabs word error rate (23.8%) is higher than the unprocessed input (17.0%).

↓ lower is better, ↑ higher is better. Dropped and changed words are per 100 reference words. Word error rate also counts words the transcriber added, so those two don’t add up to it.

Clip by clip

Of the 100 clips, Cleanvoice had fewer word errors than ElevenLabs on 52, the same number on 39 and more on 9.

52Cleanvoice betterfewer word errors
39Sameequal word errors
9ElevenLabs betterCleanvoice had more word errors

Listen: clip 0014

A damaged clip: echo, a band-reject filter and AC-3 codec damage. The speaker says: “move for the next couple of days it is not too late to get on board some have accepted it as a”. Listen for the end of the sentence, “some have accepted it as a”.

Clip 0014

Selected: Original (unprocessed)

0:00

Spectrograms of UNIVERSE clip 0014 (echo, a band-reject filter and AC-3 codec damage) through each tool, with what Omnilingual ASR heard. Cleanvoice: 5% of words wrong, keeps the end of the sentence, "some have accepted it as a". Unprocessed input: 14% wrong, the end comes out as "asam". Adobe Podcast: 27% wrong, stops at "some". Auphonic: 32% wrong. ElevenLabs and Descript Studio Sound: 36% wrong. Dolby UNIVERSE: 55% wrong. Each of them drops or garbles the end of the sentence. — Clip 0014 through six of the tools. The dashed line is 8 kHz, where the 16 kHz input ends; anything above it was restored by the tool. Transcripts by Meta’s Omnilingual ASR (CTC 7B v2).

Which one to choose

When to choose ElevenLabs

You already work in ElevenLabs. Keeping one tool can be worth more than the gap on your material; test your own files.

When to choose Cleanvoice

You want an AI voice isolator that drops fewer words. On clip 0014 (echo and codec damage), the transcriber heard "son excepte sa" at the end of the ElevenLabs output; on Cleanvoice it heard "some have accepted it as a", as in the clean original. Across 100 clips, ElevenLabs dropped 7.1 words per 100 against 3.0.
You want the voice to stay the speaker's own: speaker similarity 0.955 vs 0.928, and SCOREQ 4.54 vs 3.72.
You want enhancement that only works on the sound. Cleanvoice Studio Sound removes background noise, echo and other voices and restores bandwidth. It has no step that cuts or rewrites words.
You may also want editing later. Filler-word, stutter and silence removal are separate, optional Cleanvoice edits. You choose whether to use them, and Studio Sound works without them.
You want the same model in the web app and the API, so a manual workflow can become an automated one without changing the sound.

More on what it fixes: Cleanvoice Studio Sound, the AI audio enhancer.

How we tested

Data: the 100 noisy clips of Dolby’s public UNIVERSE validation set (Serrà et al., CC BY 4.0; noises include DEMAND, CC BY-SA).
Processing: each tool’s own product or API on 1 November 2026, at default enhancement settings. Cleanvoice processed each clip separately; the other tools processed the 100 clips joined into one 8.4-minute, 16 kHz file, which we cut back into clips. Dolby UNIVERSE is the paper’s released research outputs.
Word metrics: Meta’s Omnilingual ASR (CTC 7B v2) transcripts of each output against the same model on the clean originals (1,423 words). CTC models generally don’t hallucinate the way models with a language model do. Word error rate also counts words the transcriber added, so dropped and changed words don’t add up to it.
Quality: SCOREQ and DNSMOS P.835 OVRL. Speaker similarity, ESTOI and PESQ are measured against the clean original.
Limits: one dataset, 100 clips, 16 kHz input only, an automatic transcriber, and a test run by Cleanvoice itself.

The full method, every tool’s numbers and more audio are on the speech enhancement benchmark page.

ElevenLabs comparison questions

Still have a question?

Send it over and a real person will answer.

support@cleanvoice.ai

Ask Now

In our test, yes: fewer word errors (15.5% vs 23.8%), fewer dropped words (3.0 vs 7.1 per 100) and changed words (11.2 vs 15.3), higher SCOREQ (4.54 vs 3.72) and speaker similarity (0.955 vs 0.928), and more bandwidth (17.3 vs 11.8 kHz).

Of the 100 clips, Cleanvoice had fewer word errors on 52, the same number on 39 and more on 9.

On clip 0014, which has echo, a band-reject filter and AC-3 codec damage, the transcriber heard "son excepte sa" at the end of the ElevenLabs output, where the clean original says "some have accepted it as a": 36% of the words were wrong. On the Cleanvoice output it heard the whole ending, with 5% of the words wrong.

It has no step that cuts or rewrites words: it removes noise, echo and background voices and restores bandwidth. It can still get a word wrong on badly damaged audio, as every tool in this test did. The transcriber found 15.5 word errors per 100 on Cleanvoice's output, against 17.0 on the unprocessed input. Filler-word, stutter and silence removal are separate Cleanvoice edits you turn on yourself.

We ran the 100 noisy, 16 kHz clips of Dolby's public UNIVERSE validation set through each tool on 1 November 2026, at default settings. Cleanvoice and DeepFilterNet3 processed each clip on its own. The other products processed the 100 clips joined into one 8.4-minute file, which we cut back into clips. Dolby's row uses the outputs published with its paper. We transcribed every output with Meta's Omnilingual ASR (CTC 7B v2), a CTC model, and compared it with the same model on the clean originals (1,423 words). CTC models generally don't hallucinate the way models with a language model do. Word error rate also counts words the transcriber added, so dropped and changed words don't add up to it. SCOREQ and DNSMOS rate the sound without a reference; ESTOI and PESQ compare it with the clean original.

Cleanvoice AI Research ran it. It covers one dataset of 100 clips, mostly English VCTK speech, and 16 kHz input only. Tools ran on 1 November 2026 at default settings, and other audio or settings could give other results. The method, every number and the audio are on the speech enhancement benchmark page.

Compare

More comparisons and the full benchmark

Adobe, Auphonic, Descript, ElevenLabs and Dolby are trademarks of their owners. Tests run by Cleanvoice on 1 November 2026; results may differ on other material.